Dynamic Graph Convolution Network with Multi-head Attention for Traffic Flow Prediction

Abstract

Purposes Traffic flow prediction is crucial for the effective management and operation of urban transportation systems. The flows of different road sections or intersections in a traffic network change dynamically with time, meanwhile the flows of spatially neighboring road sections or intersections affect each other. In order to better learn the spatial and temporal correlation of the traffic flow of different road sections or intersections from the traffic flow sequences, and to improve the performance of short-term prediction of traffic flow, in this paper we propose a traffic flow prediction method based on Dynamic Graph Convolution Network with Multi-head Attention (DGCNMA). Methods The DGCNMA model first introduces graph convolution networks into the Transformer framework to learn the spatial embedding of traffic flow sequences and incorporate them into the traffic flow sequences, and then adopts the mechanism of Multi-head Attention to capture the temporal and spatial correlation of the traffic flow sequences from multiple perspectives at the same time; second, the Interactive Dynamic Graph Convolution Network is introduced to simultaneously learn the local and global spatial-temporal correlations of traffic flow sequences through the interactive learning of convolutional network and dynamic graph convolutional network, and the interactive fusion of parity subsequence features. Findings Experiments on highway traffic flow datasets (PEMS03, PEMS04, PEMS08) and subway crowd flow datasets (HZME inflow and HZME outflow) show that the proposed DGCNMA model has better traffic flow prediction performance than the baseline models.